Using Trip Information for PHEV Fuel Consumption Minimization 27 th International Battery, Hybrid and Fuel Cell Electric Vehicle Symposium (EVS27) Barcelona, Nov. 17-20, 2013 Dominik Karbowski, Vivien Smis-Michel, Valentin Vermeulen Work funded by the U.S. DOE Vehicle Technology Office (Program Managers: David Anderson, Lee Slezak)
Optimal Energy Management of xevs Needs Trip Prediction Vehicle energy use can be reduced through application of control theory or fine tuning: Dynamic Programming (DP): finds the global optimum for the command law Instantaneous optimization: o ECMS: Equivalent Minimization Consumption Strategy o PMP: Pontryagin Minimization Principle All techniques require knowledge of the trip! Increased connectivity and increased availability of data opens the door to trip prediction GPS Sensors Geographical Information mph 60 40 20 0 0 10 20 30 Miles Trip prediction Live Traffic Situation On-board Computing Cloud Computing 2
Trip Prediction 3
Modeling Vehicle Speed with Markov Chains What is a Markov chain? Collection of random variables {,,, } Memoryless: the future only depends on the present, not the past = =, =,, = )= = = )=, Homogenous, i.e. the probability, does not depend on time 17.0 Speed (m/s) 16.5 16.0 15.5 14.5 14.0 13.5 P=0.05 P=0.15 P=0.2 P=0.3 P=0.15 P=0.1 P=0.05 Vehicle speed can be represented by a Markov chain: Random variable can be vehicle speed: Speed at time t+1 only depends on speed at time t Random variable can be vehicle speed and acceleration: Speed at time t+1 depends on speed and acceleration at time t t-2 t-1 t Time (s) t+1 The Markov chain is defined by a Transition Probability Matrix (TPM): =,,,,
TPM Is Created from Real-World Trips From the CMAP Database: CMAP = Chicago Metropolitan Agency for Planning Data acquired as part of a comprehensive travel and activity surveyfor northeastern Illinois in 2007-2008 9000+ trips / 400+ drivers / 6,000,000 data points Data filtered to remove outliers and unrealistic trips 20 100 number of occurrences (%) 15 10 5 0 0 5 10 15 20 0 distance(miles) 80 60 40 20 cumulative distribution (%) 5
Constraining the Markov Chain to the Characteristics of a Given Segment Target segment defined by representative variables: Average speed Distance Speed limit Generated segment: Actual speed Average speed Distance Number of stops The Performance Value PV quantifies how close to the target the generated segment is: = + TPM Random number generation + + max(,0) Initialization (t=0, a=0, v=v 0 ) Compute next state d>d tgt? v=v end? PV<α Speed Profile Yes Yes Yes No No No 6
Example of Segment Speed Limit 50 km/h Target Speed 32 km/h 7
Combining Markov Chains and Geographical Information Itinerary in GIS (ADAS-RP) Vehicle Speed Raw Data Formatting + Segmentation Distance Speed (km/h) Synthesized Trip V V max V avg t stop 120 100 80 60 40 Iterative Stochastic Generation for each Segment for segment= 1 to n 20 0 700 Elevation (m) 650 600 550 0 10 20 30 40 Distance (km) end 8
Example of Entire Trip 90 V V tgt max V avg act V avg t stop 80 70 Speed (km/h) / Time (s) 60 50 40 30 20 10 0 0 100 200 300 400 500 600 Time (s) 9
Itinerary Used for Study on Control Munich area ~ 36 km Speed limited to 100 km/h Speed (km/h) Elevation (m) 120 100 80 60 40 20 0 700 650 600 V V max V avg t stop Same target, but 10 different synthesized trips Grade is same for all synthesized trips 550 0 5 10 15 20 25 30 35 40 Distance (km) Distance (km) 10
Optimal Control 11
Definition of the Problem Simulation environment: Autonomie, forward-looking ~ Prius 2012 PHEV: Battery: 4 kwh, 200 V, Li-ion Rated all-electric range: 26 km Top EV speed = 100 km/h Baseline Control Strategy EV Charge-Sustaining: Rule-based Optimum system efficiency look-tables ICE ON SOC + t 1 itinerary, 10 different trips Can the knowledge of the trip help reduce the fuel consumption?
Optimal Control Uses Pontryagin s Minimization Principle The high-level command variable is the battery power At each time step, the optimal command is the one that minimises the Hamiltonian: Hamiltonian =argmin ( + ( ) ) Fuel Power Function of through optimal operation maps Equivalence Factor Term close to 1 Battery Power Command In our study we make the assumption that = PMP only in Charge-depleting mode, then baseline Charge-Sustaining mode control 13
The Challenge of PMP: the Equivalence Factor Depends on the Trip! SOC too high: Electricity is too expensive There is battery energy left at the end of the trip Worst fuel consumption than baseline optimal SOC target too low: Electricity is too cheap Battery is discharged too early Missed opportunity to displace more fuel 14
Case 1: Equivalence Factor Is Optimal for each Trip Equivalence factor optimal for each trip = > best case scenario Different eq. factor for each trip Fuel savings: 3.5 to 5.7 %, 4.6% on average 15
Operations with Optimal Control After this point, control switches to CS mode 16
In Real-World, the Exact Speed Profile Is not Known! 120 100 Predicted * * Actual Speed (km/h) 80 60 40 20 0 0 5 10 15 20 Distance (km) Prediction will never match actual speed because of the stochastic nature of driving => Eq. factor will not necessarily be optimal How good is the optimization if this case? * Both trips are synthesized; Predicted and Actual labels for illustration purpose 17
Using an Equivalence Factor not Optimized for Actual Speed Profile still Brings Benefits 1 point = 1 trip, 1 eq. factor 1 shape/color = 1 trip Average benefit for a given eq. factor over all 10 trips Average benefit for a given eq. factor over top 8 trips Equivalence Factor 18
Conclusion Improving vehicle energy efficiency through connectivity requires both control optimization and trip prediction. Each has to be implementable! Trip prediction is achieved using a combination of Markov chains and GIS: A GIS (e.g.: ADAS-RP) provides trip-specific information Predicted Trip = aggregation of stochastic micro-trips that fits constraints from GIS Optimal control using trip prediction: Achieved through PMP controller Key factor for PMP efficacy, equivalence factor, depends on trip Benefits of the technology can be evaluated in simulation; in our sample itinerary (not statistically representative): best case scenario (eq. factor is adapted to trip): 4-6 % fuel savings real-world scenario (one eq. factor per itinerary): 3-4% fuel savings Future Work Trip prediction: refine process and integrate in Autonomie Optimal control: Develop fast optimal equivalence factor prediction algorithm for PMP Implement an adaptive equivalence factor Run large-scale study to quantify in a statistically representative way the benefitsof trip-based control 19
Acknowledgement HERE(formerly NAVTEQ) provided free license for ADAS-RP Work funded by the US Department of Energy Vehicle Technology Office; Program managers: David Anderson, Lee Slezak Contact Dominik Karbowski (Principal Investigator): dkarbowski@anl.gov/ 1-630-252-5362 Aymeric Rousseau (Systems Modeling and Control Manager): arousseau@anl.gov See also www.autonomie.net 20
Backup Slides 21
30 25 20 15 10 5 0 0 5000 10000 distance (m) From Itinerary Definition to Simulation of Optimal Control Define Itinerary in GIS (ADAS-RP) Generate Speed Profile Using Markov Chains Speed/Grade Compute Controller Optimal Tuning Optimal Controller speed (m/s) 55 SOC (%) 50 45 40 35 r=2.58 r=2.6 r=2.62 r=2.64 r=2.66 r=2.68 r=2.7 r 0 30 r=2.72 25 0 500 1000 1500 Time (s) speed (m/s) 25 20 15 Speed/Grade 10 5 0 0 5000 10000 distance (m) Simulate in a Forward-Looking Model (Autonomie) Our approach: Work on both optimal control and prediction Propose implementable solutions 22
Operating the Powertrain Optimally (with a few Givens) Requires Optimal Operation Maps One mode power-split offers freedom, and no-easy optimum : Engine speed can be controlled independently from vehicle speed Depending on vehicle and engine speed, there is energy recirculation (inefficient) An offline algorithm computes the optimal operating point for given output speed, torque demand and battery power out T (Nm) gb = 10 kw out T (Nm) gb 1000 500 0 0 2000 4000 1000 out ω (rpm) eng (rpm) 500 ω gb ω gb T eng (Nm) 0 0 2000 4000 out (rpm) 180 160 140 120 100 4000 3000 2000 out T (Nm) gb out T (Nm) gb 1000 500 0 0 2000 4000 1000 out ω (rpm) eng (rpm) 500 ω gb ω gb T eng (Nm) 0 0 2000 4000 out (rpm) 150 100 50 =7 kw 4000 3000 2000 23
Argonne Published Work Adaptive control: P. Sharer, A. Rousseau, et al., Plug-in Hybrid Electric Vehicle Control Strategy: Comparison between EV and Charge-Depleting Options, SAE paper 2008-01-0460, SAE World Congress, Detroit, April 2008 Dynamic Programing: D. Karbowski, A. Rousseau, et al., Plug-in Vehicle Control Strategy: From Global Optimization to Real Time Application, 22th International Electric Vehicle Symposium (EVS22), Yokohama, October 2006 D. Karbowski, K.-F. Freiherrvon Pechmann, et al., Comparison of Powertrain Configurations for Plug-In Hybrid Operation Using Global Optimization, SAE paper 2009-01-1334, SAE World Congress, Detroit, April 2009 Pagerit, S., Rousseau, A., Sharer, P., Global Optimization to Real Time Control of HEV Power Flow: Example of a Fuel Cell Hybrid Vehicle, EVS 21, Monaco, April 2005. PMP / ECMS: N. Kim and A. Rousseau, "A Sufficient Condition of Optimal Control for Hybrid Electric Vehicles,"IMechEPart D: J. Automobile Engineering, accepted. N. Kim, S. W. Cha, and H. Peng, "Optimal Equivalent Fuel Consumption for Hybrid Electric Vehicles," IEEE Trans. Control Syste. Technol., vol. 20. no. 3, May 2012, pp. 817-825. [link] N. Kim, A. Rousseau, and D. Lee, "A Jump Condition of PMP-based Control for HEVs,"J. Power Sources, vol.196, no.23, Dec. 2011, pp. 10380-10386. [link] N. Kim, S. W. Cha, and H. Peng, "Optimal Control of Hybrid Electric Vehicles Based on Pontryagin'sMinimum Principle," IEEE Trans. Control Syste. Technol., vol. 19, no. 5, Sept. 2011, pp. 1279-1287. Trip prediction: D. Karbowski, S. Pagerit, A. Calkins, "Energy Consumption Prediction of a Vehicle along a User-Specified Real- World Trip, EVS26, May 2012, Los Angeles 24